Bots made most web traffic in 2025 and AI writing reached about half of new articles; that share has stalled, detectors are unreliable, and fraud is the measured harm.
- Bots are the majority. Automated traffic passed half, but a large share is ordinary infrastructure.
- AI text has stalled. Graphite puts it near half since early 2025; a study from Imperial College London, the Internet Archive and Stanford found less.
- Detectors are signals, not proof. They struggle with edited text and wrongly flag many non-native English writers.
- Fraud is the harm with a hard number. The FBI's 2025 report gives AI-related complaints their own section, and suggests many victims do not realise AI was involved.
Every claim on this site is supposed to carry a source. This page is where the sources live. Four times a year this page gathers the newest published measurements of how much of the internet is machine-made — traffic, text, models, detection, fraud — puts them in one place, dates them, and notes what they do not prove. Numbers below are the most recent published figures as of July 2026.
① Traffic: the machines hold the majority
The thirteenth annual Bad Bot Report, published April 2026 by Imperva (a Thales company), found that automated traffic accounted for more than 53% of all web traffic in 2025, up from 51% the year before. Human activity fell to roughly 47%. This was the second consecutive year machines outnumbered people, which matters more than the first: one crossing is an anomaly, two is a trend line.Source: Imperva / Thales, 2026 Bad Bot Report, published Apr 2026, full-year 2025 data. Verbatim: “automated traffic continues to outpace human activity online, accounting for more than 53% of all web traffic in 2025, up from 51% the year before”, and “27% of bot attacks target APIs”. Re-read 17 Sep 2026.
The same edition finds that 27% of bot attacks now target APIs rather than pages: increasingly the traffic does not even pass through the human-facing internet. The full report sits behind a contact form on Imperva’s site. Trade coverage repeats further figures from it — the share of traffic that is malicious bots, and how fast AI-driven attacks grew — but they were not read at source here, so this page does not quote them.Imperva / Thales, 2026 Bad Bot Report, Apr 2026, read at source 17 Sep 2026: “27% of bot attacks target APIs”.
"Bots" is not a synonym for "fakes." Part of that automated share is legitimate: search engine crawlers, uptime monitors, accessibility tools, RSS readers, and the AI crawlers that index the web — including the ones that may have brought you here. The dead-internet claim is about who the content is for, not merely who requests it.
Do not read “bots” as “fakes”. Much automated traffic is ordinary infrastructure; the question that matters is who the content is for, not who requests it.
② Text: about half of new articles, and holding
Research firm Graphite sampled English-language articles from Common Crawl and classified them with multiple detection tools. Their May 2026 study found that primarily AI-generated articles reached 35.9% within a year of ChatGPT's launch, 48% within two years, and have held near half since the start of 2025 — 49.9% in Q1 2026.Source: Graphite, “AI Now Writes as Many Online Articles as Humans”, May 2026, read at source 17 Sep 2026: “After only 12 months, primarily AI-generated articles accounted for 35.9% of articles published.” and “In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026.”
The widely repeated prediction that 90% of online content would be synthetic by 2026 did not happen. It was repeated for years in the name of a 2022 Europol report, and Europol has since disowned the statement behind it.Europol Innovation Lab, Facing reality? Law enforcement and the challenge of deepfakes, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.” The current (January 2024) version therefore cannot be cited for the forecast.
The curve flattened. Publishing volume is not the same as consumption: Graphite also found that AI-generated articles are published in enormous quantity but appear far less often among top-ranking search results, which still skew human-written or heavily human-edited.
A second count, and it disagrees. In April 2026 researchers at Imperial College London, the Internet Archive and Stanford sampled websites first archived across 33 monthly intervals from August 2022 to May 2025, stratified by archival date, MIME type, URL depth and top-level domain, and ran detection over them. They put AI-generated or AI-assisted text at roughly 35% of newly published websites by mid-2025 — about seventeen points below Graphite’s 51.7% for May 2025.Dolezal, Alam, Graham & Bohacek, “The Impact of AI-Generated Text on the Internet”, arXiv:2604.26965, Apr 2026 (Imperial College London, Internet Archive, Stanford), read at source 17 Sep 2026: “We find that by mid-2025, roughly 35% of newly published websites were classified as AI-generated or AI-assisted, up from zero before ChatGPT's launch in late 2022.”
The gap is not a contradiction. One counts whole sites from the Internet Archive; the other counts individual articles from Common Crawl; and they run different detectors. They are answering different questions and both answers are defensible. Keeping both here is the point of this page.
That study also tested two of the things the dead internet theory actually predicts. Semantic diversity: supported — AI prevalence correlated with semantic contraction (ρ=0.47, p=0.004), with AI-written sites about a third more similar to each other than human-written ones. Factual accuracy: not supported — the correlation between AI likelihood and factual error rate was not significant (ρ=−0.19, p=0.27). Their survey found 75.1% of respondents expected the opposite on accuracy.Dolezal, Alam, Graham & Bohacek, “The Impact of AI-Generated Text on the Internet”, Apr 2026, full text read at source 23 Sep 2026: “the average semantic similarity between websites predicted to be AI-generated or AI-assisted was 33% higher than that of non-AI websites”, and “While 75.1% of respondents leaned towards agreement with this hypothesis, our quantitative analysis did not find a statistically significant correlation between the factual error rate and the aggregate AI likelihood score (ρ=−0.19, p=0.27).”
This cannot be counted cleanly. Most writing now passes through a human and a model — outlined by one, drafted by the other, edited back. Detectors classify a spectrum as a binary. Read every "X% of the web is AI" figure, including the ones on this page, as an estimate produced by imperfect detectors, not a census.
Quote a “share of the web is AI” figure as an estimate, with its method. Two careful counts differ because they measure different things with different detectors.
③ Detection: signals, not proof
The tools that produce those percentages are weaker than their marketing. The RAID benchmark (University of Pennsylvania, ACL 2024) evaluated twelve detectors against 6.2 million generations across eleven models, eight domains and eleven adversarial attacks, finding substantial difficulty generalising to unseen models and domains and steep accuracy drops from simple changes such as synonym swaps or a repetition penalty. Stanford research published in Patterns in July 2023 tested seven widely used detectors on 91 TOEFL essays written by non-native English speakers and found an average false-positive rate of 61.3%, while essays by US eighth-graders were classified with near-perfect accuracy.Sources: Dugan et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors”, read at source 11 Sep 2026 — “over 6 million generations spanning 11 models, 8 domains, 11 adversarial attacks and 4 decoding strategies”. ACL 2024 (6.2m generations, 11 models, 8 domains, 11 adversarial attacks; 12 detectors evaluated) · Liang, Yuksekgonul, Mao, Wu & Zou, “GPT detectors are biased against non-native English writers”, Patterns 4(7):100779, July 2023.
Detector makers claim very high accuracy — “99% or more”, as the RAID authors summarise the claims — but a headline accuracy figure says little about how a detector handles text that has been changed: in RAID, adversarial attacks and repetition penalties were enough to fool current detectors.Dugan et al., RAID, ACL 2024, abstract read at source 17 Sep 2026: “Many commercial and open-source models claim to detect machine-generated text with extremely high accuracy (99% or more).” and current detectors “are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models.” This is why the site's guide to spotting fake media treats detection results as signals, not proof, and why no one should face consequences on a detector score alone.
Do not act on a detector score by itself. Headline accuracy comes from clean test material, and independent tests found detectors weakest on edited text and on non-native English writing.
④ Models: the shipping rate
One public tracker counts 249 models released since ChatGPT's launch in November 2022, and calls them all frontier models — a count that depends entirely on where a tracker draws that line. This page's July 2026 edition gave 198, recorded then from the same tracker and lab announcements; that earlier snapshot can no longer be checked. Below the frontier labs sits a wave of open-weights releases that anyone can download.Source: AI Release Tracker, read 17 Sep 2026: “We cover 249 tracked frontier models from Anthropic, OpenAI, Google, Meta, SpaceXAI, DeepSeek, Mistral, Moonshot AI, Z.ai, Qwen, NVIDIA, starting with the launch of ChatGPT on November 30 2022.” The count grows with every release; re-query it. See who builds AI.
⑤ Fraud: AI gets its own section
The FBI's 2025 Internet Crime Report, from the Internet Crime Complaint Center (IC3), gives artificial intelligence a section of its own: 22,364 complaints and $893,346,472 in losses where the information reported referred to AI. The 2024 report had no such section. Within that total, investment complaints with a reported AI nexus came to just over $632M; business email compromise involving AI more than $30M; AI-involved employment scams almost $13M.Source: FBI, 2025 IC3 Annual Report, read at source 17 Sep 2026: “In 2025, IC3 received more than 22,000 complaints reporting AI-related information. Adjusted losses of these complaints exceed $893 million.” and “In 2025, losses in Investment complaints with a reported AI-nexus, surpassed $632 million.”
Total reported losses reached $20.877 billion across 1,008,597 complaints, and people aged 60 and over reported roughly $7.7 billion — the largest losses of any age group, up 59% on 2024. The report also suggests the AI figure understates the problem: investment scams as a whole lost more than $8 billion, which it reads as a sign that many victims do not realise AI was involved. A cloned voice that worked leaves no evidence that it was cloned.FBI, 2025 IC3 Annual Report, read at source 17 Sep 2026: “1,008,597 complaints; $20.877 billion in losses” · “60+: 201,266 complaints, $7.7 billion in losses.” · elder-fraud section: “$7.748 Billion in Losses”, “59% FROM 2024” · “However, overall losses to Investment scams exceeded $8 billion, demonstrating that many victims do not realize the extent AI may be involved in scams.”
Treat the AI fraud figure as a floor, not a total. It counts only complaints where AI was mentioned, and across all reported losses, people over 60 lost the most.
The five numbers, in one table
| Measure | Figure | Period | Source (date published) | Direction |
|---|---|---|---|---|
| Automated share of web traffic | 53% | Full-year 2025 | Imperva / Thales Bad Bot Report (Apr 2026) | Rising |
| New articles primarily AI-generated | ~50% | Q1 2026 | Graphite (May 2026) | Flat since Q1 2025 |
| Losses in US complaints mentioning AI | $893M | Full-year 2025 | FBI IC3 Annual Report (2025 edition) | New section |
| Models released since ChatGPT, one tracker's count | 249 | Nov 2022 – Sep 2026 | AI Release Tracker (read Sep 2026) | Rising |
| Detector false-positive rate, non-native English writers | >60% | 2023 study | Liang et al., Patterns (2023) | Unresolved |
Reasoning, September 2026 — this table restates figures from the sections above; each one is linked and quoted from its source there.
Method, and what this page is not
The full sourcing rules live on the method page; errors are corrected in place on this page, with a note where the change matters.
- Every figure comes from a named, published source with a date printed beside it. Where a number is contested, the dispute is stated.
- Primary sources are preferred: the report or paper itself, or first-hand reporting of it. Vendor marketing claims and SEO roundups are excluded.
- Nothing here is original measurement. This site does not crawl the web; this page gathers, dates, and contextualizes the work of others.
- Figures are snapshots. Sources publish on their own schedules — annually, continuously, or irregularly — and this page is re-cut every quarter around whatever is newest.
- Superseded numbers move to the archive below rather than disappearing, so the trend line stays visible.
The measurements are real and the direction is uncomfortable — but the honest reading of this quarter is not that the internet ended. Automated traffic passed half; a large share of it is ordinary infrastructure. AI writing reached about half of new articles and then stopped climbing, and it still loses to human work where readers actually are. Detection remains too weak to accuse anyone. The genuine harm with a hard number attached is fraud, and in dollars it lands hardest on people over 60. Worry is warranted. Hostility toward the technology is not the same thing as attention to the evidence.
Archive
Q3 2026 — this edition. First publication. Headline figures: bots 53% of traffic (2025), AI-written articles ~50% (Q1 2026), AI fraud losses $893M (2025), 198 frontier models since Nov 2022. Revised 17 Sep 2026: the model count now reads 249 on the tracker, the Graphite share is dated to its May 2026 study, and the fraud figures are quoted from the FBI report itself.First-hand: this is the July 2026 edition’s own headline list, kept as first published; the current figures are linked and quoted in the sections above.
Q4 2026 — due October 2026.
Sources
- Imperva / Thales — 2026 Bad Bot Report: Bots in the Agentic Age, published April 2026 (full-year 2025 data).
- Graphite — AI Now Writes as Many Online Articles as Humans, May 2026, superseding its October 2025 study.
- Dolezal, Alam, Graham & Bohacek — The Impact of AI-Generated Text on the Internet, arXiv, April 2026 (Imperial College London, Internet Archive, Stanford).
- Europol Innovation Lab — Facing reality? Law enforcement and the challenge of deepfakes, 2022, revised January 2024.
- FBI Internet Crime Complaint Center — 2025 IC3 Annual Report, and the 2024 edition for comparison.
- RAID benchmark — University of Pennsylvania, ACL 2024.
- Liang et al. — GPT detectors are biased against non-native English writers, Patterns, 2023.
- AI Release Tracker — model release count, read September 2026.